Causal, Econometric and Marketing-Mix Models

Introduction

Demand forecasting becomes strategically valuable when it explains not only what is likely to happen, but also why demand changes and how managers can influence it. A time-series model may predict sales accurately from historical patterns, yet provide little guidance about the likely effect of a price reduction, advertising campaign, competitor launch, recession, or major event. Causal and econometric models address this limitation by linking demand to economically meaningful drivers.

Marketing-mix models are an important application of econometric demand analysis. They estimate how sales respond to price, distribution, promotion, advertising, competitors, seasonality, macroeconomic conditions, and other interventions. Their purpose is not merely to fit historical data. A credible model should support decisions such as:

• What is the expected sales effect of a 5% price reduction?
• How much incremental demand was generated by a promotion?
• Which advertising channel has the highest return on investment?
• How rapidly does advertising influence sales, and how long does the effect persist?
• Would a competitor’s price increase create an opportunity for the firm?
• What sales would have occurred without a campaign?

These questions are causal. They require careful attention to identification, confounding, reverse causality, multicollinearity, measurement, and model validation.

Learning Objectives

By the end of this chapter, students should be able to:

1. Distinguish predictive forecasting from causal demand estimation.
2. Specify and interpret regression-based demand models.
3. Construct distributed-lag, dynamic, and marketing-mix models.
4. Evaluate price, promotion, advertising, competitor, macroeconomic, and event effects.
5. Diagnose multicollinearity, autocorrelation, heteroskedasticity, omitted-variable bias, and endogeneity.
6. Explain identification strategies, including experiments, instrumental variables, fixed effects, and difference-in-differences.
7. Assess model validity, profitability, managerial usefulness, and ethical risks.

1. From Correlation to Causal Demand Analysis

A forecasting relationship describes statistical association. A causal relationship asks what would happen to demand if a variable were deliberately changed while relevant conditions were held appropriately constant. The causal effect of a price change is therefore not automatically equal to the observed correlation between price and sales.

Suppose a retailer lowers price when it expects weak demand. Historical data may show that low prices occur alongside low sales. A naive regression could conclude that lower prices reduce demand, which contradicts economic theory. The problem is that expected demand affects the firm’s pricing decision. Price is endogenous: it is related to unobserved determinants of demand.

A useful conceptual distinction is:

• Descriptive model: summarizes observed relationships.
• Predictive model: forecasts future outcomes under expected conditions.
• Causal model: estimates the outcome of an intervention or counterfactual action.

A model can be highly predictive without being causally valid. For example, a holiday indicator may forecast sales effectively, but it does not imply that changing the calendar would increase demand.

Potential outcomes and counterfactuals

Let Y_i(1) denote demand for market unit i if a treatment, such as a promotion, occurs, and let Y_i(0) denote demand if it does not occur. The individual causal effect is:

Only one of these potential outcomes is observed for a given unit at a given time. The causal-inference problem is therefore fundamentally a missing-counterfactual problem. Regression, experiments, and quasi-experimental methods attempt to estimate an average treatment effect under explicit assumptions.

In demand analysis, the “treatment” may be continuous rather than binary. Price, advertising expenditure, and distribution coverage can take many values. The analyst then estimates a demand response or elasticity rather than a simple treated-versus-untreated effect.

2. The Structural Demand Model

A general demand equation can be written as:

where:

• Q_t = quantity demanded in period t, measured in units, packages, or revenue;
• P_t = own price;
• A_t = advertising or communication activity;
• M_t = promotion variables, such as displays, coupons, or discounts;
• C_t = competitor variables, such as competitor prices or advertising;
• X_t = macroeconomic and market conditions;
• Z_t = seasonality, holidays, events, distribution, and other controls;
• ε_t = unobserved demand shocks.

The function f may be linear, logarithmic, nonlinear, or specified through a generalized linear model. A simple linear model is:

Here, β₀ is the intercept and each coefficient represents the expected change in demand associated with a one-unit change in the relevant explanatory variable, conditional on the other included variables.

The phrase “conditional on” is essential. The coefficient on price is not necessarily a causal price effect unless the model includes appropriate controls and the price variation is plausibly exogenous or otherwise identified.

Functional forms and elasticity

A log-log model is commonly used when proportional responses are more meaningful than absolute responses:

Under this specification, β_P is the price elasticity of demand. It measures the approximate percentage change in quantity associated with a 1% change in price, holding other included variables constant. If β_P = −1.4, a 1% price increase is associated with an approximate 1.4% decrease in quantity.

A log-level model, such as ln(Q_t) = β₀ + β_A A_t + ε_t, interprets β_A as the approximate proportional change in demand associated with a one-unit increase in advertising expenditure. A level-log model interprets the coefficient as a unit change in quantity associated with a 1% change in the predictor. The analyst must state the functional form and units clearly.

Revenue and profit implications

Demand estimation should ultimately support economic decisions. Revenue is:

If marginal cost is MC, profit is:

where F represents fixed costs. For a constant-elasticity demand curve, the markup condition for an interior profit maximum is:

where ε_P is the own-price elasticity and must be less than −1 for the standard positive-markup solution. This result assumes a single product, known marginal cost, no capacity constraint, a differentiable demand curve, and a price chosen by the firm. Multi-product firms require consideration of cross-price effects and substitution.

3. Building a Marketing-Mix Model

A marketing-mix model estimates the contribution of controllable marketing variables while accounting for nonmarketing influences. A practical baseline specification is:

For panel data, i may identify a store, region, product, or customer segment, while t identifies a time period. The terms α_i are unit fixed effects and δ_t are time fixed effects.

A typical variable architecture is as follows:

Promotions

Promotion variables may be binary indicators, continuous discount depth, or categorical treatments. A simple model may include separate indicators for a feature, display, and coupon. However, these activities may occur together, so the individual coefficients are identified only when there is sufficient independent variation.

A discount can be represented as:

The definition must be consistent across observations. If regular price itself changes, a discount measure based on a contaminated reference price may misrepresent the treatment.

Advertising carryover and adstock

Advertising often affects demand beyond the period in which it is purchased. A geometric adstock transformation is:

where A_t^* is effective advertising stock, A_t is current advertising input, and λ is a retention parameter satisfying 0 ≤ λ < 1. If λ = 0.5, half of the previous effective stock carries forward each period. This transformation assumes a constant geometric decay and should not be treated as a universal law. The retention parameter may be estimated, calibrated from research, or varied in sensitivity analysis.

Advertising may also exhibit diminishing returns. One specification is:

The logarithm requires a nonnegative advertising measure and implies that the incremental effect decreases as advertising stock increases. Alternative saturation functions include Hill, logistic, and negative exponential forms. Choice should be guided by theory, data range, and out-of-sample performance rather than fit alone.

Interactions

Marketing effects may depend on context. For example:

The marginal effect of advertising is:

Thus, β_A alone is not the advertising effect at every price. Centering variables can improve interpretability and reduce nonessential correlation between main effects and interaction terms, although centering does not solve fundamental identification problems.

4. Distributed-Lag and Dynamic Demand Models

Immediate-response models can be unrealistic. Consumers may delay purchase, advertising may build awareness gradually, and promotions may cause pantry loading followed by a post-promotion decline.

A finite distributed-lag model is:

The impact effect is β₀, while the cumulative effect over K periods is:

The coefficients can be difficult to estimate separately because adjacent lags are often highly correlated. Polynomial distributed lags, Almon restrictions, or geometric decay impose a more parsimonious structure.

An autoregressive distributed-lag model includes lagged demand:

If the model is stable, meaning |ρ| < 1, the long-run effect of a permanent one-unit increase in X is:

This formula applies to the specified linear model and assumes a constant long-run relationship, stable parameters, and appropriate treatment of other dynamic variables. A high estimated ρ may represent genuine persistence, but it may also reflect omitted trends or nonstationarity.

Stationarity and spurious regression

A stationary series has stable probabilistic properties such as mean and variance over time. Regressing one trending variable on another can produce a high R² and apparently significant coefficients even when no meaningful relationship exists. Analysts should examine plots, unit-root diagnostics, economic theory, first differences, growth rates, or cointegration methods where appropriate.

Differencing a variable is defined as:

Differencing can remove persistent trends but may also discard long-run information. Cointegration analysis is appropriate only when theory and statistical evidence support a stable long-run equilibrium relationship among nonstationary variables.

5. Identification and Causal Validity

Identification concerns whether the available variation in a predictor can isolate its causal effect. In a regression model:

ordinary least squares requires, among other conditions, a zero conditional mean assumption:

This means that, after conditioning on included variables, the explanatory variables are not systematically related to unobserved determinants of the outcome. If the condition fails, the coefficient may be biased and inconsistent.

Omitted-variable bias

Suppose the true model is:

but W is omitted and correlated with price. The estimated price coefficient absorbs part of the effect of W. For example, a retailer may raise price when a product receives unusually strong demand from a news event. Without controlling for the event, the price coefficient may be biased upward.

Controls should be selected because they are confounders or otherwise needed for the estimand, not simply because they improve in-sample fit. Controlling for a variable caused by the treatment can create post-treatment bias.

Reverse causality and simultaneity

Price and quantity may be jointly determined. Firms may set price in response to expected demand, while demand responds to price. Competitor price may also respond to the firm’s price. These are simultaneity problems, not merely ordinary measurement errors.

Experiments

Randomized price tests, geo-experiments, controlled promotions, and advertising holdouts provide strong identification when implementation is credible. Randomization makes treatment assignment independent of potential outcomes in expectation. Practical concerns include interference between markets, noncompliance, limited duration, novelty effects, and external validity.

Fixed effects

Panel models with unit fixed effects control for time-invariant differences across stores, regions, or products. Time fixed effects control for shocks common to all units in a period:

Fixed effects do not control for time-varying omitted variables, nor do they automatically solve price endogeneity. Standard errors should generally account for within-unit serial correlation, often through clustering at the treatment-assignment or relevant geographic level.

Difference-in-differences

If one group receives a campaign and another comparable group does not, a difference-in-differences estimator is:

where T denotes treatment and C control. The central assumption is parallel trends: absent treatment, the average change in the treatment group would have matched the average change in the control group. Analysts should inspect pre-treatment trends and consider staggered adoption, anticipation, spillovers, and treatment-effect heterogeneity.

Instrumental variables

An instrument Z can identify the causal effect of an endogenous variable X if it satisfies:

1. Relevance: Z is correlated with X.
2. Exclusion: Z affects demand only through X.
3. Independence: Z is not related to unobserved demand shocks.

The exclusion and independence assumptions are substantive and cannot be established solely by a strong first-stage statistic. Weak instruments produce imprecise and potentially misleading estimates. An instrument may identify a local average treatment effect for units whose treatment responds to the instrument, not necessarily a universal effect.

6. Multicollinearity and Statistical Diagnostics

Multicollinearity occurs when predictors contain overlapping information. Price, discount, display, advertising, and seasonality may all move together. Consequences include large standard errors, unstable coefficients, sensitivity to specification, and difficulty separating individual contributions. Multicollinearity does not necessarily bias coefficients, but it reduces precision and interpretability.

The variance inflation factor for predictor j is:

where R_j² is obtained by regressing predictor j on the other predictors. A large VIF is a warning signal, not a universal proof of an invalid model. Remedies include better experimental variation, combining substantively similar measures, regularization for prediction, restrictions based on theory, longer panels, or redesigning the data collection process. Removing a variable solely because it is collinear can create omitted-variable bias.

Other diagnostics include:

• residual plots for nonlinearity and unequal variance;
• autocorrelation functions and tests for serial dependence;
• influence and leverage diagnostics;
• out-of-sample and rolling-origin validation;
• placebo tests and pre-trend tests for quasi-experiments;
• sensitivity analysis over functional forms, lag lengths, and control sets;
• tests for structural breaks after major market changes.

Heteroskedasticity affects conventional standard errors. Robust or clustered standard errors may improve inference, but they do not repair endogeneity or misspecification. Autocorrelation requires time-aware inference and sometimes explicit dynamic modeling.

7. Measurement, Data Architecture, and Model Development

A credible marketing-mix system begins with a clearly defined estimand. “Effect of advertising” could mean short-term incremental sales, total sales including carryover, profit contribution, or long-run brand equity. These are different outcomes.

A disciplined workflow is:

1. Define the decision, outcome, treatment, unit, time horizon, and estimand.
2. Develop a causal diagram or written causal theory.
3. Assemble price, promotion, media, distribution, competitor, macroeconomic, event, and inventory data.
4. Align timing and units; distinguish planned from delivered media and list from transaction price.
5. Explore trends, seasonality, missingness, outliers, and data revisions.
6. Estimate a theoretically defensible baseline model.
7. Diagnose residuals, collinearity, dynamics, and sensitivity to specification.
8. Validate using holdout periods or geographically separated data.
9. Translate effects into incremental volume, revenue, margin, and return on investment.
10. Monitor drift and re-estimate when market behavior changes.

A marketing-mix coefficient should not be interpreted as incremental contribution if the model fails to represent baseline demand, distribution constraints, inventory stockouts, or campaign selection. Sales may be observed only when products are available; stockouts can make demand appear lower than it truly is.

Incrementality and return on investment

If a campaign produces estimated incremental quantity ∆Q, contribution margin per unit is m, and campaign cost is C, incremental profit is:

Return on marketing investment can be defined as:

The organization must state whether m is gross margin, contribution margin, or another profitability measure. Using revenue instead of contribution may exaggerate economic value.

8. Illustrative Case: FreshBite Ready Meals

Background

FreshBite, a fictional ready-meal brand, sells through 600 supermarkets. Management observes that weekly unit sales rise during feature promotions and periods of television advertising. However, promotions are more frequent in stores with weak baseline sales, and price reductions are often introduced shortly before holidays. Management wants to estimate incremental demand and determine whether advertising should be expanded.

Problem

A simple comparison of promoted and nonpromoted weeks shows that promoted weeks have only 8% higher sales. Managers argue that the promotion is ineffective. The analytics team suspects confounding because promoted stores have lower baseline demand and promotions are concentrated in holiday periods.

Analysis and approach

The team constructs a store-week panel. The dependent variable is ln weekly units sold, with controls for transaction price, display, coupon, distribution, stockout rate, holiday indicators, competitor price, local unemployment, store fixed effects, and week fixed effects. Television expenditure is transformed using a geometric adstock with retention parameter λ = 0.6, while the value of λ is varied in sensitivity analysis.

The baseline model is:

The team also conducts a four-week geo-experiment. Twenty comparable markets are randomly assigned to receive additional television support, while twenty markets serve as controls. Pre-treatment sales trends are examined, and standard errors are clustered by market.

The observational model estimates a larger promotion effect after controlling for holidays, store differences, and stockouts. The experiment produces a smaller but positive advertising effect. The discrepancy leads management to conclude that the original observational estimate partly captured favorable holiday timing and selection of stores into campaigns.

Decision and outcome

FreshBite adopts three policies:

1. Use randomized geographic tests for major advertising reallocations.
2. Use the panel model for forecasting and scenario analysis, but report sensitivity ranges rather than a single point estimate.
3. Evaluate campaigns using incremental contribution margin rather than sales revenue alone.

The firm shifts some budget from broad television exposure to markets where the estimated marginal response remains high and distribution is sufficient. It also records stockouts separately so that future models distinguish unavailable demand from weak consumer interest.

Learning points

The case illustrates five principles. First, raw promoted-versus-unpromoted comparisons are not causal estimates. Second, fixed effects and time controls reduce some confounding but do not guarantee exogeneity. Third, advertising carryover must be modeled explicitly. Fourth, experiments are valuable for validating observational models. Fifth, the managerial objective is incremental profitable demand, not merely statistical significance.

9. Managerial, Ethical, and Governance Considerations

Causal models influence prices, media allocation, promotions, and customer treatment. Managers should therefore document data provenance, model versions, assumptions, estimands, and uncertainty. A model that recommends higher prices in areas with low purchasing power may be economically rational under a narrow objective but socially controversial or legally restricted.

Potential governance risks include discriminatory price personalization, use of sensitive personal data, opaque automated decisions, excessive surveillance, and misleading attribution claims. Privacy laws and competition regulations vary by jurisdiction. Analysts should minimize data collection, use appropriate aggregation or anonymization, control access, and obtain legal review when personal or sensitive data are involved.

AI and machine learning can improve feature construction, nonlinear prediction, and media optimization, but predictive importance is not equivalent to causal influence. Black-box models should not be used to claim that a channel caused sales growth unless identification is addressed. A strong practice is to combine machine learning for nuisance prediction or heterogeneous-effect discovery with experimental or quasi-experimental research designs.

10. Contemporary Developments and Best Practices

Current practice increasingly combines structured econometric models with experimentation and machine learning. Important developments include:

• geo-randomized media experiments and synthetic controls;
• causal forests and heterogeneous treatment-effect analysis;
• Bayesian hierarchical models for partial pooling across products and markets;
• state-space models for changing coefficients and latent demand;
• privacy-preserving measurement and aggregated media data;
• multi-touch attribution models calibrated against incrementality tests;
• scenario systems that propagate parameter uncertainty into price and budget decisions.

Best practice is not to select the most complex model. It is to select the simplest model that represents the decision, data-generating process, and causal assumptions adequately. Every reported effect should include a definition, unit, time horizon, uncertainty measure, identification strategy, and sensitivity assessment.

Chapter Summary

Causal, econometric, and marketing-mix models connect demand to economically meaningful drivers such as price, promotion, advertising, competition, distribution, macroeconomic conditions, and events. Regression models provide a flexible foundation, but causal interpretation requires more than a statistically significant coefficient. Analysts must define the estimand, specify a credible counterfactual, address omitted variables and reverse causality, and assess whether the available variation identifies the desired effect.

Distributed-lag and dynamic models capture delayed response, persistence, advertising carryover, and long-run effects. Panel fixed effects, experiments, difference-in-differences, and instrumental variables can strengthen identification when their assumptions are plausible. Multicollinearity, serial correlation, heteroskedasticity, nonstationarity, measurement error, stockouts, and structural breaks can undermine inference. Finally, marketing decisions should be evaluated using incremental contribution and uncertainty, while respecting privacy, fairness, transparency, and regulatory requirements.

Key Terms

Causal effect: The change in an outcome attributable to an intervention relative to its counterfactual.

Counterfactual: The outcome that would have occurred under an alternative action or condition.

Econometric model: A statistical model grounded in economic theory and used to estimate relationships among variables.

Endogeneity: A condition in which an explanatory variable is related to unobserved determinants of the outcome.

Exogeneity: A condition in which explanatory variables are sufficiently unrelated to the model error for the intended inference.

Marketing-mix model: A model estimating the effects of marketing and environmental variables on demand or sales.

Price elasticity: The percentage response of quantity demanded to a 1% change in price.

Distributed lag: A model in which current and past values of a predictor affect the outcome.

Adstock: A transformation representing the carryover of advertising effects over time.

Fixed effects: Controls for unobserved, time-invariant differences across units.

Difference-in-differences: A method comparing changes over time between treated and comparison groups.

Instrumental variable: A variable used to identify an endogenous causal effect under relevance and validity assumptions.

Multicollinearity: Strong correlation or overlapping information among explanatory variables.

Heteroskedasticity: Nonconstant variance of the model error.

Autocorrelation: Dependence among errors across time or ordered observations.

Stationarity: Stability of a series’ probabilistic properties over time.

Incrementality: The additional outcome caused by a marketing action.

ROMI: Return on marketing investment, commonly incremental contribution net of cost divided by cost.

Identification: The conditions under which a causal parameter can be recovered from observed data.

Review Questions

Short Answer Questions

1. How does a causal demand model differ from a predictive demand model?
2. What does a price elasticity of −1.4 mean in a log-log demand model?
3. Why can advertising require an adstock transformation?
4. What problem does a unit fixed effect address?
5. State the relevance and exclusion conditions for an instrumental variable.

Long Answer Questions

1. Explain how omitted-variable bias and reverse causality can distort an estimated price coefficient. Use a retail example.
2. Compare finite distributed-lag models with autoregressive distributed-lag models, including their interpretation of short-run and long-run effects.
3. Discuss the assumptions, strengths, and limitations of difference-in-differences for evaluating a promotion.
4. Explain how multicollinearity affects marketing-mix analysis and distinguish remedies for prediction from remedies for causal identification.
5. Design a complete workflow for estimating and validating a marketing-mix model for a national consumer brand.

Analytical / Application Questions

1. A model estimates Q_t = 1,000 − 20P_t + 5A_t, where price is measured in dollars and advertising in thousands of dollars. Explain why the coefficient on price should not automatically be interpreted causally, and identify two design changes that could improve identification.
2. A retailer reports that promoted weeks have higher sales, but promotions are concentrated in holidays and low-baseline stores. Propose a panel-data or experimental design that could estimate incremental promotion effects.
3. A campaign generates 12,000 incremental units, contribution margin is $3 per unit, and campaign cost is $20,000. Calculate incremental profit and ROMI, then identify two reasons why the estimate may still be uncertain.

Multiple Choice Questions with Answers

Question 1

In a log-log demand model, the coefficient on own price is interpreted as:

A. The dollar change in revenue from a one-unit price change

B. The approximate percentage change in quantity from a 1% price change

C. The percentage change in price from a one-unit quantity change

D. The probability that a customer purchases the product

Correct Answer: B. The approximate percentage change in quantity from a 1% price change

Question 2

Which condition is most directly required for ordinary least squares to provide an unbiased causal coefficient?

A. The dependent variable must be measured in dollars

B. All predictors must be normally distributed

C. The explanatory variable must be unrelated to the error conditional on included controls

D. The model must contain no intercept

Correct Answer: C. The explanatory variable must be unrelated to the error conditional on included controls

Question 3

In the adstock equation A_t^* = A_t + λA_t−1^*, what does λ represent?

A. The price elasticity of demand

B. The advertising retention or carryover parameter

C. The campaign’s fixed cost

D. The competitor’s market share

Correct Answer: B. The advertising retention or carryover parameter

Question 4

A major consequence of multicollinearity is:

A. Automatically biased coefficients in every regression

B. Larger standard errors and unstable individual coefficient estimates

C. Guaranteed improvement in out-of-sample accuracy

D. Elimination of omitted-variable bias

Correct Answer: B. Larger standard errors and unstable individual coefficient estimates

Question 5

The parallel-trends assumption is central to which method?

A. Difference-in-differences

B. Geometric adstock

C. Ordinary seasonal decomposition

D. Simple moving averages

Correct Answer: A. Difference-in-differences

Question 6

For the dynamic model Q_t = α + ρQ_t−1 + βX_t + ε_t, the long-run effect of a permanent one-unit increase in X is:

A. β(1−ρ)

B. β/ρ

C. β/(1−ρ) when |ρ|<1

D. ρ/β

Correct Answer: C. β/(1−ρ) when |ρ|<1

Question 7

Which statement about an instrumental variable is correct?

A. It must affect the outcome through several independent pathways

B. It must be correlated with the endogenous predictor and satisfy a credible exclusion restriction

C. It can be selected solely because it improves R²

D. It removes the need to define a causal estimand

Correct Answer: B. It must be correlated with the endogenous predictor and satisfy a credible exclusion restriction

Question 8

If incremental contribution is $36,000 and campaign cost is $20,000, ROMI defined as net incremental contribution divided by cost is:

A. 0.80

B. 1.20

C. 1.80

D. 2.80

Correct Answer: B. 1.20

Question 9

Which issue can cause observed sales to understate true consumer demand?

A. A stockout that prevents purchase

B. A positive contribution margin

C. A randomized treatment assignment

D. A correctly measured holiday indicator

Correct Answer: A. A stockout that prevents purchase

Question 10

Why should predictive machine-learning importance not automatically be interpreted as causal marketing contribution?

A. Machine-learning models cannot process numerical data

B. Predictive association may reflect confounding rather than intervention effects

C. Causal models never use historical data

D. Marketing variables are always normally distributed

Correct Answer: B. Predictive association may reflect confounding rather than intervention effects